arkennemasis fal · Flux 3 Image to Video · $0.17-0.29/s

Image-to-video is the honest way to use a video model. Text-to-video asks the model to invent a whole shot; image-to-video asks it to add time to a picture you already chose - so the composition, the lighting and the character are settled before you spend a cent. You keep the frame you like and pay only for the motion.

ArkFal_blackforestlabs_flux_3_image_to_video animates a single still into a clip with synchronized audio, up to 20 seconds, at $0.17/s (720p) or $0.29/s (1080p). It's FLUX 3 Video on fal, which as of now is the only place FLUX 3 Video runs - flux-2.md records it as early-access API with private weights and the open Dev backbone still unshipped.

What you set

• image - required. The still the clip starts from. PNG, JPEG or WebP; the node uploads it to fal's CDN for free before the paid call.
• prompt - what should happen. Not a description of the picture; a description of the motion. "She turns to camera, hair moves, camera slowly pushes in" beats a paragraph about how pretty she is.
• resolution - 720p (default) or 1080p. The price difference is real: 1080p is about 70% more per second.
• duration - auto, or pin the seconds. Video models are seconds-billed; this box is your budget.
• generate_audio - on by default, and free either way. Dialogue, SFX and ambience arrive with the clip, which is a big part of why FLUX 3 feels different from the silent open models.
• aspect_ratio - auto lets the model choose. With an image input that's usually sensible; set it explicitly if the still's shape and your target don't agree.
• safety_tolerance - 0 strictest to 4 most permissive, default 2.

Shared optional controls: max_cost_usd (default $20; stops the run before upload if the estimate exceeds it - for 1080p work a $5 cap is not paranoid), reuse_identical_run (on by default), extra_json, max_concurrent.

Outputs: video, video_path, seed (so a good roll can at least be noted down), info.

The draft habit

The companion ArkFal_blackforestlabs_flux_3_image_to_video_draft does the same job at $0.06/s and returns a draft cache you can finish at full quality through draft_enhance. If you're going to run the same still three ways - which is exactly what you'll do - draft all three, pick, and finish one. Ten seconds of 1080p full render is $2.90; ten seconds of draft is $0.60.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis

Then restart. (ComfyUI Manager → search arkennemasis → install does the same thing with the dependency install handled.) Your key goes in .env:

FAL_KEY=your-key-from-fal.ai/dashboard/keys

next to run_nvidia_gpu.bat, or in the ComfyUI folder on a regular install. There's no key input on the node - a widget would end up baked into every workflow JSON you share.

Where people get burned

Check the still's resolution before you connect it. The model works to a target resolution you select, so a 512px source stretched to 1080p will look soft by construction - upscale first or accept the softness. And if the source has alpha, don't expect it downstream: alpha is dropped on the way through the pipeline in both directions.

The other one is expectation-setting. Independent reviewers of the FLUX 3 showcase have already caught the characteristic failure mode - physics that look right at a glance and wrong on the second look. Add time, and things that were impossible in a still become visible. Keep shots short, keep motion motivated, and it holds up far better.

Files go to output/fal/blackforestlabs_flux-3_image-to-video/, and every paid request is logged in output/fal/_requests.jsonl - the node's recovery net if ComfyUI dies mid-render.